Last year, a grieving air traveler asked Air Canada’s chatbot about bereavement fares. The bot invented a refund policy that did not exist. The customer acted on it, the airline ended up in court, and the story went viral. The court rejected Air Canada’s argument that its chatbot was a “separate legal entity” responsible for its own actions and ordered the airline to pay damages.

That incident is now a cautionary tale for every brand scaling AI in customer communications. And new research from customer communications platform Sinch suggests it is far from a one-off.

Some 74% of enterprises have already been forced to roll back a deployed AI agent due to governance failures, according to Sinch’s “AI Production Paradox” report. Here is the twist: companies with the most mature guardrails, those that invested most heavily in compliance, safety protocols, and oversight, rolled back at an even higher rate of 81%. The teams doing the most to prevent failure are failing more often, not less.

“If governance was the fix, the most mature teams would roll back less, not more,” said Daniel Morris, chief product officer at Sinch. “Engineering teams are spending most of their time building and maintaining safety systems instead of focusing on improving the customer experience. That’s the guardrail tax that slows organizations down.”

The impact of the guardrail tax

For marketing teams, that guardrail tax has a direct cost. Every hour engineering spends rebuilding safety infrastructure is an hour not spent on the customer experience improvements that drive revenue.

Air Canada is not alone. A car dealership’s chatbot agreed to sell a Chevy Tahoe for $1 after a prank prompt. An AI support bot at the coding startup Cursor invented a nonexistent login policy, triggering a wave of customer cancellations. A delivery company’s bot swore at a customer and wrote a poem trashing its own employer. Each incident went viral. Each damaged a brand. And each one helps explain the Sinch finding that three out of four enterprises have already rolled back a deployed AI agent.

Sinch surveyed 2,527 enterprise decision-makers across 10 countries and six industries. The findings that matter most for marketers:

• 62% of enterprises already have AI communications agents in production, and 88% expect to deploy one within 12 months. The pressure to deploy is intense.
• 74% have been forced to roll back a deployed agent due to governance failures. Three out of four marketing organizations have already felt the pain of an AI rollout that had to be undone.
• 84% of teams spend at least half their engineering time rebuilding safety infrastructure from scratch. That is engineering capacity that could be going toward personalization, channel expansion, and campaign optimization.
• When an AI agent fails, 35% of the impact lands on the support queue. Nearly as much, 34%, lands on brand perception — and that one is harder to repair.  

Infrastructure quality was the single strongest predictor of deployment success, the study found, outweighing model choice, team size, and budget. Yet most organizations say their current provider falls short in at least one critical area.

AI customer communications agents handle customer conversations at scale: chatbots on websites, voice agents in contact centers, automated SMS and email responders, and omnichannel platforms that route and respond across channels. They range from simple FAQ bots to sophisticated agents that authenticate users, process transactions, and personalize responses based on customer history.

Sinch’s research tracks agents already in production, not pilots or internal experiments. These are systems marketers rely on every day, where a failure means frustrated customers, longer wait times, and brand damage that spreads in minutes.

Picking the wrong foundation is the real risk

Jayashree Iyangar, global lead of CX data and AI at HGS, a digital experience firm, said the findings match what she sees in the field. Marketers are past the pilot phase, she noted, and the real challenge lies in operations.

“The key question is how AI can be orchestrated seamlessly across multiple channels, not whether it can be deployed in one,” Iyangar said.

She pointed out that the risk profile varies significantly by use case. A marketing chatbot that fumbles a promotional offer carries less weight than a service agent that mishandles a sensitive billing issue. “Human-in-the-loop oversight remains central in service environments where the risk of negative customer impact is higher,” she said. “That’s also where we see more instances of AI rollbacks.”

Her take on infrastructure echoes Sinch’s core finding. “A significant portion of effort is being spent on building and maintaining safety systems rather than advancing customer experience,” she said. She is seeing organizations consolidate around centralized AI governance teams that handle trust, compliance, and security separately from the AI use cases themselves.

Three moves marketers can make now

For marketing teams, the study points to three practical moves.

1. Let infrastructure drive your vendor decision. Infrastructure quality predicts deployment success more than any other variable in the Sinch data. When evaluating providers, ask about guardrail engineering, cross-channel orchestration, and the extent to which your team will absorb the safety burden. The right platform handles most of the safety work, so your team can focus on customer experience.
2. Plan for the guardrail tax in your roadmap. Safety systems are not a one-time setup cost. They consume ongoing engineering resources that would otherwise be devoted to CX improvements. Budget for that reality from the start rather than watching your timeline slip when rollbacks hit.
3. Push for a separate governance function. Iyangar’s observation about centralized AI governance teams aligns directly with the data. Keeping AI use cases and governance engineering separate reduces overhead. Marketing should not own safety infrastructure. It should partner with a dedicated governance function that handles trust, compliance, and security, freeing marketing to focus on work that directly touches customers.

Source link